CoolFace
Datasetpublic

thientrangngv/SERA-KimiK3-Django-SWEAgent-Raw-T1

SERA Kimi-K3 Django SWE-Agent - RAW T1 (first rollout) 300 raw, pre-postprocess first-rollout agent trajectories generated with the SERA SVG pipeline (paper), using Kimi K3 as the teacher. Released raw so you can choose your own filtering, verification threshold and reasoning-trace handling. Companion: SERA-KimiK3-Django-SWEAgent-Raw-T2. Schema Mirrors allenai/Sera-*-T1/T2: column notes messages JSON string - apply json.loads(). Raw SWE-agent history:… See the full description on the dataset page: https://huggingface.co/datasets/thientrangngv/SERA-KimiK3-Django-SWEAgent-Raw-T1.

sourceHugging Faceodc-byupdated 1mo agoView on Hugging Face
1likes38downloads
Dataset Card

SERA Kimi-K3 Django SWE-Agent - RAW T1 (first rollout)

300 raw, pre-postprocess first-rollout agent trajectories generated with the SERA SVG pipeline (paper), using Kimi K3 as the teacher.

Released raw so you can choose your own filtering, verification threshold and reasoning-trace handling. Companion: SERA-KimiK3-Django-SWEAgent-Raw-T2.

Schema

Mirrors allenai/Sera-*-T1/T2:

columnnotes
messagesJSON string - apply json.loads(). Raw SWE-agent history: role, content, thought, action, tool_calls, and thinking_blocks where the model reasoned
rollout_patchpatch produced by this rollout
target_patchT2 only: the first-rollout patch, i.e. the soft-verification reference
func_name, func_paththe sampled starting function
problem_statementT1: the vague bug prompt. T2: the synthetic PR issue
docker_imageSWE-bench container
exit_status, is_good_patch, num_agent_stepsfiltering aids (is_good_patch = the T1 LLM-grader verdict)

Nothing is filtered: every rollout is included, successful or not, so you can apply your own thresholds.

Reasoning traces

Kimi K3 returns reasoning in a channel separate from its visible text. OpenRouter exposes it as reasoning_content; it is stored here inside thinking_blocks, so content (the short outward-facing line) and the model's actual deliberation stay distinct.

  • —100% of trajectories contain reasoning
  • —~17-22% of individual assistant turns carry a reasoning block - K3 deliberates on hard steps and goes straight to a tool call on routine ones
  • —Length is extremely skewed: median ~60 chars, p90 ~600-830, but individual blocks reach 50k-114k chars

Caveat if you wrap these into `<think>` tags. Because most turns carry no reasoning and often no visible text (just a tool call), a naive add_think wrapper produces empty <think></think> on ~45-48% of turns, which teaches a student to skip reasoning. Either drop the tags on those turns or backfill from action. For reference, GLM-5.2-generated SERA data shows ~24-30% empty under the same treatment, so K3 is noticeably sparser.

Soft verification

r is recomputable at any threshold from rollout_patch vs target_patch (line-level recall), so no regeneration is needed to change it. No unit tests were used at any point, per the SERA method.

Generation config

Teacher Kimi K3 (moonshotai/kimi-k3 via OpenRouter), harness SWE-agent (str_replace_editor, bash, submit), SWE-bench Django container django__django-7530 (base f8fab6f9), 115-step cap, temperature 0.6 (SERA default), SERA soft-verified generation (SVG).

Generation cost ~$900 for 300 first-rollout and 160 second-rollout trajectories. The second stage was stopped by a spend guard partway through, which is why T2 has fewer rows than T1.